Trust-Aware Deep Learning Framework for Mitigating False Data Injection and Integrity Breaches in IoT Ecosystems

Abstract

IoT systems grow explosively so they are becoming more vulnerable to False data injection (FDI) and data integrity attacks. This can impact the availability and reliability of the system. Standard security mechanisms are insufficient to provide the appropriate security level required for large, dynamic, heterogeneous environments. In this paper we propose a Trust Aware Deep Learning Framework (TADLF) for intelligent detection and prevention of integrity, related cyber-attacks for IoT. We propose a Dynamic Trust Assessment (DTA) component in order to calculate trust coefficient scores by determining device trustworthiness using historical data transmission behavior, source authenticity, and device communication records. In addition, an attention, based hybrid CNN BiLSTM module is proposed to effectively learn the spatiotemporal anomalies. By combining the DTA and an extended feature set extracted based on the trust score, the attention, based hybrid model is achieved average accuracy of 98.4%, precision of 98.1%, recall of 98.6% and F1 score of 98.3% with 1.5% false positive alarm rate, IoT dataset having 461,000 data records, for FDI attacks compared to base models (RF, XGBoost and LSTM).